Most neuromorphic chips take their architecture from the cerebral cortex, the brain region that thinks. A team at Northwestern University instead built a chip from a different part of the brain entirely. The cerebellum, the fist-sized structure at the back of the skull, does not reason so much as predict, constantly comparing what the body is about to do with what it expects to happen, and its computational signature is not amplification but restraint. The chip that emulates it, reported July 10 in Nature Communications, runs the same way: it stays quiet for the expected and spends its energy only on the unexpected.
The hardware is built from molybdenum disulfide memtransistors, devices that combine memory and transistor action in a single atomically thin channel, eliminating the constant shuttling of data between separate memory and processing units that slows conventional computers. The Northwestern devices add a twist: asymmetry. One electrode contacts the semiconductor directly while the other sits partly above it, separated by a thin insulating extension. Reversing the voltage polarity reconfigures the same device between two modes that the team calls excitatory and inhibitory, mirroring the competing signals that keep the cerebellum’s circuits in balance.
Why the balance matters
In the cerebellum, excitatory and inhibitory signals normally cancel each other. Granule cells excite Purkinje cells, the large output neurons famous for the strongest short-term facilitation in the brain, while molecular-layer interneurons deliver feed-forward inhibition. When a familiar input arrives, the two signals stay in equilibrium and the circuit responds predictably. When something novel appears, the balance shifts and the network flags the change.
The memtransistors reproduce that dance in hardware. In the excitatory configuration the devices show a gradual, slow-decaying response, analogous to the build-up of facilitation at the granule-to-Purkinje synapse. In the inhibitory configuration they respond abruptly and fade quickly, mimicking feed-forward inhibition. A novel input pushes the array out of balance, and that transient suppression is the signal the system reads as an anomaly. The team, led by materials scientist Mark Hersam, neurobiologist Indira Raman, and electrical engineer Amit Trivedi of the University of Illinois Chicago, calls the emergent behavior differentiation, the same word neuroscientists use for the Purkinje cell’s ability to distinguish input patterns.
The numbers
The performance claims come with a caveat about how they were measured. In simulated electrocardiogram tests, the system flagged arrhythmias with greater than 98 percent accuracy, compared with 70 to 80 percent for classical baseline methods, and detected anomalies within one-fifth of a heartbeat. It reached 100 percent accuracy 2.4 times faster than a transformer model, the standard architecture behind modern language and image AI. And it needed roughly 10,000 times fewer operations than silicon-based approaches to do it.
The ECG results come from network simulations built from measured single-device data, not from a fully integrated chip running on a patient’s chest. The 10,000-fold figure counts operations, not measured energy consumption; the energy savings are inferred rather than directly benchmarked. Hersam is explicit about the gap between demonstration and product: the team has not scaled the memtransistors to the level of commercial silicon chips. Device yield was around 70 percent, channels are a few hundred nanometers long, and the demonstration array is a 10-by-1 crossbar, a far cry from the billions of transistors on a modern processor.
What the cerebellum buys you
The claim worth taking seriously is not that this chip outperforms every AI system, but that it does a specific job with a fraction of the machinery. Always-on sensing, the kind of continuous monitoring a wearable heart monitor or an autonomous vehicle needs, is brutally inefficient on conventional hardware, which must process every frame, every beat, every sample, in case something matters. A cerebellum-style architecture inverts the logic: the routine is cheap because the hardware is designed to suppress it, and only the anomalous moment, the arrhythmia, the pedestrian stepping into the road, triggers a full response.
That is the same design principle the cerebellum applies to motor control, where the brain must react to perturbations within tens of milliseconds without the latency of conscious thought. The Northwestern team’s next target is habituation, the cerebellum’s ability to learn what counts as routine in the first place, which would let the chip adapt its definition of normal as it watches the world. The group has also shown the approach generalizes beyond electrocardiograms, demonstrating novelty detection on handwritten digits and spoken audio.
The work joins a broader shift in edge computing, where power constraints, not raw intelligence, are the binding limit. Data centers already consume hundreds of terawatt-hours of electricity per year, and running a large model on every wearable is not an option. Chips that borrow the cerebellum’s economy, that sleep through the boring parts, are one plausible path out of that bind. The brain’s most ancient circuit, it turns out, had the right idea about attention all along: the scarce resource is not the ability to react, but the ability to know when not to.
Sources
1. Min-A Kang, Spencer T. Brown, Nethmi Jayasinghe, et al., “Cerebellum-inspired memtransistors enable emergent differentiation for hardware-efficient novelty detection,” Nature Communications (2026), published July 10, 2026. DOI: 10.1038/s41467-026-75212-4. PMID: 42431950. https://www.nature.com/articles/s41467-026-75212-4
2. Owen Hughes, “New AI chip mimics the human brain’s capacity for split-second motor control… 10,000 times fewer calculations,” Live Science, August 11, 2026. https://www.livescience.com/technology/electronics/new-ai-chip-mimics-the-human-brains-capacity-for-split-second-motor-control-it-solved-problems-using-10-000-times-fewer-calculations
3. Amanda Morris, “AI gets a cerebellum,” Northwestern Now, July 10, 2026. https://news.northwestern.edu/stories/2026/07/ai-gets-a-cerebellum
4. Semiconductor Engineering, “Research Bits: July 14,” 2026. https://semiengineering.com/research-bits-july-14/
5. Neuroscience News, “AI cerebellum memtransistor,” July 10, 2026. https://neurosciencenews.com/ai-cerebellum-memtransistor-31036
6. PubMed record: https://pubmed.ncbi.nlm.nih.gov/42431950/